[CASSANDRA-4182] multithreaded compaction very slow with large single data file and a few tiny data files - ASF JIRA

Agile Board

Attach files

Attach Screenshot

Bulk Copy Attachments

Bulk Move Attachments

Voters

Watch issue

Watchers

Create sub-task

Convert to sub-task

Move

Link

Clone

Labels

Update Comment Author

Replace String in Comment

Update Comment Visibility

Delete Comments

XML

Word

Printable

JSON

Details

Type: Improvement
Status: Resolved
Priority: Low
Resolution: Duplicate
Fix Version/s: None
Component/s: None
Labels:
None
Environment:

Redhat
Sun JDK 1.6.0_20-b02

Description

Turning on multithreaded compaction makes compaction time take nearly twice as long in our environment, which includes a very large SStable and a few smaller ones, relative to either 0.8.x with MT turned off or 1.0.x with MT turned off.

compaction_throughput_mb_per_sec is set to 0.

We currently compact about 500 GB of data nightly due to overwrites. (LevelDB will probably be enabled on the busy CFs once 1.0.x is rolled out completely) The time it takes to do the compaction is:

451m13.284s (multithreaded)
273m58.740s (multihtreaded disabled)

Our nodes run on SSDs and therefore have a high read and write rate available to them. The primary CF they're compacting right now, with most of the data, is localized to a very large file (~300+GB) and a few tiny files (1-10GB) since the CF has become far less active.

I would expect the multithreaded compaction to be no worse than the single threaded compaction, or perhaps a higher cost in CPU for the same performance, but it's half the speed with the same CPU usage, or more CPU.

I have two graphs available from testing 2 or 3 compactions which demonstrate some interesting characteristics. 1.0.9 was installed on the 21st with MT turned on. Prior stuff is 0.8.7 with MT turned off, but 1.0.9 with MT turned off seems to perform as well as 0.8.7.

http://www.xney.com/temp/cass-irq.png (interrupts)

http://www.xney.com/temp/cass-iostat.png (io bandwidth of disks)

This demonstrates a large increase in rescheduling interrupts and only half the bandwidth used on the disks. I suspect this is because some kind of threads are thrashing or something like that.